PaperlayerView exact revision on arXiv

Source-linked AI summary

MOLA LiDAR-Inertial Odometry (MOLA-LIO) on the COMFORT Localization Benchmark

Jose Luis Blanco-Claraco

arXiv:2609.25813v1cs.RO

TL;DR

With test-phase references hidden when configuration work began, the report selects a single MOLA-LIO configuration using local held-out missions and reference classes. Its benchmark results show strong relative translation error but higher absolute error than the best entry, while the discussion identifies vertical drift as an unresolved boundary.

  • Problem

    Test-phase references were hidden when configuration work began, so configuration decisions had to be made locally rather than against those references.

    Source: §IV · p. 2

  • Method

    MOLA-LIO combines scan-to-map LiDAR odometry with IMU fusion in a causal, single-pass system, selecting one configuration through local held-out checks.

    Source: A. Architecture · p. 1, B. From LO to LIO · p. 1, §IV · p. 2

  • Results

    RTE=0.0098 m is within 9% of the best entry, while ATE is 2.4 times the best entry’s.

    Source: §VI · p. 2

  • Takeaways & Limitations

    The submitted smoother was chosen because it reports an honest pose covariance, whereas the lightweight filter reports a constant floor.

    Source: §IV · p. 2

    The framework lacks an anchor on verticality able to bound drift over a ten-minute mission.

    Source: §VI · p. 2

Abstract

from arXiv · show

This short report documents our entry to the COMFORT Localization Benchmark (IROS 2026), evaluated on the GrandTour dataset recorded with the Boxi payload. It extends MOLA-LO into a LiDAR-inertial system that also ingests IMU and, optionally, legged kinematic odometry. We describe the architecture, the streams consumed, the local protocol that selected the submitted configuration, and the measurements backing our real-time claim.

I. REPORT SUMMARY

Table I consolidates the declarations required by the benchmark organizers.

A. Architecture

The architecture combines reusable YAML-configured blocks for time-ordered input, LiDAR scan registration, and state estimation; the submitted system runs online in one pass.

A. Architecture

  • MOLA configures a SLAM system as reusable blocks declared in YAML, without writing code.Source: A. Architecture · p. 1
  • The input stage merges a mission’s per-topic ROS 1 bags into one time-ordered stream.Source: A. Architecture · p. 1
  • The front end registers each scan to the local map using mp2p_icp generalized-ICP with point-wise covariances.Source: A. Architecture · p. 1
  • The submitted configuration uses a fixed-lag IMU-preintegration smoother over a 1.0 s window and runs single-pass, without offline re-optimization or map post-processing.NavStateFilter supplies the ICP initial guess; the alternative is a lightweight SE(3) filter.Source: A. Architecture · p. 1

B. From LO to LIO

Relative to LO, MOLA-LIO adds inertial fusion and covariance-to-covariance matching, while asynchronous map rebuilding reduces insertion time and spikes.

B. From LO to LIO

  • The IMU levels the initial pose, constrains gravity direction, and contributes preintegrated factors between keyframes in the smoother.Source: B. From LO to LIO · p. 1
  • Cov-to-cov matching forms and weights pairings from local covariances of both clouds rather than point geometry alone.Source: B. From LO to LIO · p. 1
  • Background k-d tree rebalancing cuts local-map insertion time from 81.0 s to 7.2 s over a 436 s mission and removes insertion spikes.Source: B. From LO to LIO · p. 1
  • Legged kinematic odometry was implemented as a full SE(3) observation with loose velocity sigmas, but did not improve held-out scores and is disabled.Source: B. From LO to LIO · p. 1

III. SENSOR MODALITIES AND TOPICS

The system uses selected Boxi-payload sensors and reports poses in the benchmark’s required frame; resampling improves association under the scorer’s time window at a small RMS cost.

III. SENSOR MODALITIES AND TOPICS

  • The Hesai LiDAR won screening on all missions against the other two payload LiDARs, and the ADIS16475 was preferred over the higher-rate STIM320.Source: §III · p. 1
  • The scans use the dataset’s points_undistorted stream, already motioncompensated from the robot’s legged kinematic-inertial state estimate.Source: §III · p. 1
  • Reported poses use the body-frame base estimate right-composed with the static base-to-prism transform and resampled to 200 Hz; cameras and GNSS/INS products are not used.Table II lists the consumed GrandTour bag streams.Source: §III · p. 1, Table II · p. 1
  • Under the scorer’s 10 ms window, resampling raises heap-1 associations from 868 of 4598 reference samples to 4576, at a cost of 1.8 mm RMS.The passage characterizes resampling as an association device, not a source of information.Source: §III · p. 2

IV. HOW THE SUBMITTED CONFIGURATION WAS CHOSEN

Configuration choices were made locally on held-out missions after test references became hidden, using reference classes ordered by trust. One frozen configuration was submitted for all six test missions; on four held-out missions it reached 2.10 cm ATE, without separating the two state estimators.

IV. HOW THE SUBMITTED CONFIGURATION WAS CHOSEN

  • Configuration decisions used held-out missions and three reference classes ordered by trust: total-station prism, qualified GNSS/INS proxy, then reference-free checks.The prism trace is the benchmark scoring reference; GNSS/INS was used only when its error was well below the screened effect, otherwise checks chiefly compared attitude with the legged state estimator.Source: §IV · p. 2
  • 2.10 cm ATE against the prism reference on four held-out missions; 59% of the error was vertical.The screen did not separate the smoother (2.10) from the lightweight filter (2.09); the smoother was submitted for its honest pose covariance, while the filter reports a constant floor.Source: §IV · p. 2
  • One frozen configuration was uploaded for all six test missions, with no per-mission tuning.Legged odometry fused as an absolute pose cost 1.1 mm against a 0.1 mm run-to-run floor, so it was disabled.Source: §IV · p. 2

V. RUNTIME AND REAL-TIME OPERATION

The submitted system runs causally in a single pass, and runtime varies across missions, with con-4 determining the real-time margin. Local-map insertion creates the slowest scans, while registration dominates average per-scan processing; asynchronous rebuilding also trades repeatability for speed.

V. RUNTIME AND REAL-TIME OPERATION

  • The submitted configuration is causal and single-pass, with Table III reporting wall-clock time divided by mission duration alongside per-mission score and runtime costs.Measurements were taken one run at a time on an idle machine, with the solver free to use all 16 hardware threads; latency is per scan at a 100 ms scan period.Source: §V · p. 2, Table III · p. 2
  • Con-4 is the mission the real-time claim turns on; its radius-adaptive point budget triples cloud sizes.The slowest scan takes 2 to 5.4 scan periods, with local-map insertion accounting for 452 of con-4’s 544 ms; the resulting short backlog is absorbed rather than dropped.Source: §V · p. 2
  • ICP accounts for 47–67% of the per-scan pipeline, and correspondence formation and weighting account for 81–86% of ICP.The smoother accounts for 5–11% and map insertion for 2–6% of the per-scan pipeline.Source: §V · p. 2
  • The pipeline reaches only 7.0 of 16 hardware threads, with serialization—not throughput—as the bound.Local-map insertion drives the slowest scans, whereas registration dominates average per-scan time.Source: §V · p. 2
  • The asynchronous local-map rebuild makes repeated runs differ by 0.7 to 15.5 mm RMS after alignment.Synchronous rebuilding brings repeat agreement to about 1 mm at roughly 30% more wall clock; the faster setting was reported because the deterministic configuration missed con-4 runtime.Source: §V · p. 2

VI. DISCUSSION

The discussion contrasts strong relative translation error with weaker absolute error and identifies unbounded vertical drift over a ten-minute mission as future work. It also argues that six-mission averages can be highly sensitive to one difficult sequence.

VI. DISCUSSION

  • RTE=0.0098 m is within 9% of the best entry, while ATE is 2.4 times the best; 77–86% of squared error is vertical on construction-site missions.The discussion identifies the missing element as an anchor on verticality able to bound drift over a ten-minute mission.Source: §VI · p. 2
  • The post-alignment residual is a smooth, low-frequency deformation, mostly vertical; an offline pose-graph correction diagnostic orders missions as the scores do.The authors identify improving the framework’s verticality anchor as future work.Source: §VI · p. 2
  • Con-4 alone contributes 44% of the six-mission average, and near-identical configurations vary by more than the gap between leaderboard positions.The authors argue that publishing per-mission ranks would make rankings less sensitive to a single hard sequence.Source: §VI · p. 2
Loading 2609.25813v1…